DeepSeek-V4-Flash-0731 on 2× NVIDIA DGX Spark (GB10, 128 GB unified) — vLLM TP2 Deployment

2-node DGX Spark deployment for DeepSeek-V4-Flash-Vision-Exp running vLLM TP=2 with DSpark speculative decoding and nvfp4_ds_mla KV cache.

Quantização Official FP8 Safetensors
Repositório deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Contexto máximo verificado1.048.576
Quantidade de dispositivos2 × NVIDIA DGX Spark (GB10, 128 GB unified)

Valor informado https://github.com/MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark

75.4 tok/s (decode-tok-s)
104.9 tok/s (decode-tok-s)
164.5 tok/s (decode-tok-s)
191.2 tok/s (decode-tok-s)

This deployment recipe is currently maintained in English because commands and technical instructions are source-aligned.

DeepSeek-V4-Flash-0731 on 2× DGX Spark (DSpark Speculative) is a working community deployment for DeepSeek-V4-Flash-0731 running on 2 × NVIDIA DGX Spark (GB10, 128 GB unified) using vLLM and Official FP8 Safetensors. Verified up to 1,048,576 tokens context.

Measured by upstream maintainerworking

Tested Setup & Specifications

Model CheckpointDeepSeek-V4-Flash-0731
Hardware SKUNVIDIA DGX Spark (GB10, 128 GB unified)
Device Count2 devices
InterconnectConnectX-7 RoCE (PTP link)
Runtime EnginevLLM (0.1.1)
Patches RequiredCustom kernel patch
Artifact FormatSAFETENSORS
Active Weights157.00 GiB
Tensor ParallelismTP = 2
Placement2 nodes, 1 GPU/node (GB10)
Verified Context1,048,576 tokens
Last Verified2026-09-23

Tested & Reported Capabilities

✓Text generation verified
✓Clean model loading & initialisation
✓Context scaling verified up to 1,048,576 tokens
✓Function & Tool calling (Verified)
✓Multimodal vision input (Verified)
✓Multi-Token Prediction / Speculative draft (Verified)

Context & Verification Notes:

1,048,576 tokens = server ceiling + full-context natural completion probe verified
899,994 tokens = explicit long-prompt acceptance measurement

Source-Backed Benchmarks

Measured performance on actual physical machines recorded by the upstream maintainer or community report. No linear extrapolations or simulated numbers.

Metric	Measured Value	Context	Concurrency	Details & Source
Decode Throughput	75.4 tok/s	—	1 stream	256 prompt, ~512 completionprompt: 256 tokSingle stream decode, TTFT 0.63 s [source ↗]
Decode Throughput	104.9 tok/s	—	2 streams	256 prompt, ~512 completion (aggregate)prompt: 256 tok2 streams aggregate (58.3 tok/s per stream), TTFT 0.81 s [source ↗]
Decode Throughput	164.5 tok/s	—	4 streams	256 prompt, ~512 completion (aggregate)prompt: 256 tok4 streams aggregate (46.8 tok/s per stream), TTFT 1.26 s [source ↗]
Decode Throughput	191.2 tok/s	—	6 streams	256 prompt, ~512 completion (aggregate)prompt: 256 tok6 streams aggregate (36.9 tok/s per stream), TTFT 1.42 s [source ↗]

Quick Start & Deployment Commands

Condensed deployment procedure extracted directly from the upstream repository.

Prerequisites

2× NVIDIA DGX Spark with ConnectX-7 point-to-point RoCE
Docker with NVIDIA Container Toolkit
Hugging Face token for gated models if using abliteration direction

Review these commands, then run them yourself. This page does not run them. This is not the deployment brief.

1Configure environment

cp .env.dspark.example .env.dspark

2Pull image on both nodes

docker pull ghcr.io/anemll/dspark-vllm-gx10:0.1.1

3Download weights on head

./prepare-dspark-model-cache.sh --official

4Launch 2-node cluster

./start-deepseek-v4-flash-dspark.sh

Caveats & Operational Constraints

Open Calculator With This Exact Setup

Source & Provenance

Mia's AI Lab: MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark — Commit: 97e8733238f81f5fdc44b241f8996a7858825744
Artifact: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp — Official FP8 Safetensors